How MT26 will be graded
MT26 locks its forecast hours before the first state's polls close, publishes how it will be graded, and grades itself on fixed dates, whatever the result.
What's locked. The public forecast file as served after the last data refresh on Tue 3 Nov, with its fingerprint and an outside copy on the Wayback Machine. Fingerprint (sha256) and Wayback copy: added here at the lock, on Tue 3 Nov.
Where results come from. AP race calls, read from AP's public results pages, or an official results page where AP hasn't called. Never the model's own "Projected" label. Each row keeps its source and the time it was read.
What's graded
| Set | Races | Use |
|---|---|---|
| Close races | The 106 rated Likely or closer by the raters at the lock (76 House, 15 Senate, 15 governor) | The headline |
| Six named races | Maine Senate, TX-15, Kansas governor, NY-17, KY-6, Alaska Senate | Each one shown |
| The 14-race group | OH-9, NY-17, CA-22, Maine Senate, Michigan Senate, PA-1, NC-1, MI-7, SC-1, TX-15, Kansas governor, AL-2, KY-6, Alaska Senate | Shown as posted on Mon 2 Nov |
| House and Senate control | Both chambers | Shown as decided or open |
| All 506 | Every race on the site | Shown in the file, never the headline: safe seats make any score look better than it is |
Comments posted on Reddit on Mon 2 Nov quoted that day's numbers. The tables show those next to the locked numbers. The locked number is the one graded.
What's reported
- Side right or wrong: did the side the model had ahead win?
- Brier score: (1 − the winner's chance)², averaged across races. Lower is better. 0 is perfect, and always saying 50 in 100 scores 0.25.
- Misses expected against actual. The model's own numbers expect misses: about 18 of the 106 on Monday's numbers. A forecast with no misses at these numbers would itself be a problem.
- Calibration bins: of the races where the side ahead was at 50-59, 60-69, 70-79, 80-89 and 90-100 in 100, how many did it win? One night proves little. These races share one national environment and tend to miss together, so one election's bins can look very good or very bad by chance.
- Direction of error: whether results ran above or below the model for the candidates on each side of the race card, not a claim about any party.
- The raters' final labels: on the same races, how often the side each rater had ahead won, with Toss-ups counted separately.
Dates
| Paris | ET | What |
|---|---|---|
| Wed 4 Nov 08:30 | 02:30 | First table: six races and both chambers, decided or open |
| Wed 4 Nov 18:00 | 12:00 | Fuller table: "graded N of 106 so far" |
| Thu 5 Nov 10:00 | 04:00 | Final table for this round, with the file for download |
Not every race will be decided by then. The headline always states its count. Maine and Alaska use ranked choice, California counts slowly, and Georgia can go to a runoff on 1 Dec. Open races stay in the table as open, with the date expected. They aren't dropped. The races called first tend to be the less close ones, so early scores flatter the model. A closing update follows on Wed 2 Dec, the day after Georgia's runoff date.
What won't be claimed. No "most accurate" claim from one night. The grade stays up, good or bad.
A forecast, not a result. MT26 is AI-built (Claude Code) by one developer and isn't affiliated with any campaign, party or outlet.